Example of predicting NPS score from feedback comments, also means of bucketizing customer sentiment.

sample-nps-nlp-model.py

import os
import pandas as pd
from sklearn.model_selection import train_test_split

import nltk
from nltk.corpus import stopwords
from nltk.tokenize import RegexpTokenizer

import time

start_time = time.time()
print('Starting run of analytic model...')

data = pd.read_csv(os.path.join(os.path.dirname(__file__), '../datasets/medallia.csv'))
def bucketize_score(x):
    if x <= 5:
        return 'Negative'
    if x == 6 or x == 7:
        return 'Neutral'
    else:
        return 'Positive'
data['NPS_score'] = data['NPS_score'].apply(lambda x: bucketize_score(x))

train, test = train_test_split(data, test_size=0.15)

comments = []
stopwords_set = set(stopwords.words('english')) # You need to download this via nltk.download(), see readme
tokenizer = RegexpTokenizer(r'\w+') # Strip punctuation and special characters from words

def tokenize_words(strng):
    strng = strng.replace('!', ' !').replace('?', ' ?').replace('.', ' .') # So that punctuation can be tokenized
    words_filtered = [e.lower() for e in tokenizer.tokenize(strng) if len(e) >= 3]
    words_cleaned = [word for word in words_filtered
        if 'http' not in word
        and word not in stopwords_set
    ]
    return words_cleaned

for index, row in train.iterrows():
    comments.append((tokenize_words(row.NPS_comment), row.NPS_score))

test_pos = test[test['NPS_score'] == 'Positive']
test_pos = test_pos['NPS_comment']
test_neu = test[test['NPS_score'] == 'Neutral']
test_neu = test_neu['NPS_comment']
test_neg = test[test['NPS_score'] == 'Negative']
test_neg = test_neg['NPS_comment']

# Extracting word features
def get_words_in_comments(comment):
    all = []
    for (words, sentiment) in comment:
        all.extend(words)
    return all

def get_word_features(wordlist):
    wordlist = nltk.FreqDist(wordlist)
    features = wordlist.keys()
    return features

w_features = get_word_features(get_words_in_comments(comments))

def extract_features(document):
    document_words = set(document)
    features = {}
    for word in w_features:
        features['contains(%s)' % word] = (word in document_words)
    return features

training_set = nltk.classify.apply_features(extract_features, comments)
classifier = nltk.NaiveBayesClassifier.train(training_set)

neg_cnt = 0
neu_cnt = 0
pos_cnt = 0
all_cnt = 0
for comment in test_neg:
    res =  classifier.classify(extract_features(tokenize_words(comment)))
    if(res == 'Negative'):
        neg_cnt = neg_cnt + 1
        all_cnt = all_cnt + 1
for comment in test_neu:
    res =  classifier.classify(extract_features(tokenize_words(comment)))
    if(res == 'Neutral'):
        neu_cnt = neu_cnt + 1
        all_cnt = all_cnt + 1
for comment in test_pos:
    res =  classifier.classify(extract_features(tokenize_words(comment)))
    if(res == 'Positive'):
        pos_cnt = pos_cnt + 1
        all_cnt = all_cnt + 1

neg_pct = (float(neg_cnt) / float(len(test_neg))) * 100
neu_pct = (float(neu_cnt) / float(len(test_neu))) * 100
pos_pct = (float(pos_cnt) / float(len(test_pos))) * 100
all_pct = (float(all_cnt) / float(len(test_pos) + len(test_neg))) * 100
print('Accuracy of model predicting bucketized NPS Score from comments:')
print('[Negative]: %s/%s (%s%%)'  % (neg_cnt, len(test_neg), neg_pct))
print('[Neutral]: %s/%s (%s%%)'  % (neu_cnt, len(test_neu), neu_pct))
print('[Positive]: %s/%s  (%s%%)' % (pos_cnt, len(test_pos), pos_pct))
print('[Overall]: %s/%s  (%s%%)'  % (all_cnt, len(test_pos) + len(test_neg), all_pct))

print('Ending run of analytic model.')
print('Runtime was %s seconds' % (time.time() - start_time))